specialized A64FX Dslash kernels

Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h

@@ -30,7 +30,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 #pragma once
 
-#if defined(DSLASHA64FX)
+#if defined(A64FX)
+
+#pragma message("invoking A64FX Dslash")
 
 // undefine everything
 #include <simd/Fujitsu_A64FX_undef.h>
@@ -194,9 +196,9 @@ WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFie
 						int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
 #include <qcd/action/fermion/implementation/WilsonKernelsAsmBody.h>
 
-#undef MAYBEPERM
+//#undef MAYBEPERM
 //#undef MULT_2SPIN
-#define MAYBEPERM(A,B)
+//#define MAYBEPERM(A,B)
 //#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
 
 /////////////////////////////////////////////////////////////////
@@ -346,7 +348,7 @@ WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, Double
 //#undef MAYBEPERM
 //#undef MULT_2SPIN
 
-// undefine everything
+// undefine 
 #include <simd/Fujitsu_A64FX_undef.h>
 
 ///////////////////////////////////////////////////////////
@@ -359,7 +361,7 @@ WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, Double
 #include <simd/Fujitsu_A64FX_asm_double.h>
 #endif
 
-// KNL stuff
+// former KNL 
 //#define MAYBEPERM(A,perm) if (perm) { A ; }
 //#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
 //#define COMPLEX_SIGNS(isigns) vComplexD *isigns = &signsD[0];

Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonImplDF/WilsonKernelsInstantiationWilsonImplDF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonImplFH/WilsonKernelsInstantiationWilsonImplFH.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc.master

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/ZWilsonImplDF/WilsonKernelsInstantiationZWilsonImplDF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/ZWilsonImplFH/WilsonKernelsInstantiationZWilsonImplFH.cc

@@ -34,9 +34,11 @@ directory
 
 #ifndef AVX512
 #ifndef QPX
+#ifndef A64FX
 #include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
 #endif
 #endif
+#endif
 
 NAMESPACE_BEGIN(Grid);
 

Grid/util/version.cc

@@ -1,5 +1,5 @@
 #include <iostream>
-#include "Version.h"
+//#include "Version.h"
 namespace Grid {
   void printHash(){
 #ifdef GITHASH